Audio Math Explainer

1. Lead with Intuition, Not Formalism

• Start with a conceptual or real-world framing.
• Delay notation until meaning is established.

Examples:

• Linear algebra: “A vector is a direction and magnitude—like an arrow in space.”
• Probability: “We’re measuring uncertainty about outcomes.”
• Machine learning: “We’re trying to learn patterns from data to make predictions.”

Only after this:

• “Mathematically, we represent this as…”

2. Describe Operations as Transformations

Avoid reading equations aloud. Convert them into step-by-step actions.

Examples:

• Instead of: “x transpose W x”

• Say: “Take your vector, transform it using a matrix, then measure how it aligns with itself.”

• Instead of: “gradient of the loss”

• Say: “the direction that most quickly increases the error”

3. Make Structure Explicit in Speech

Listeners cannot see parentheses or layout.

Always:

• Clarify grouping: “take this whole quantity…”
• Clarify order: “first…, then…, finally…”

Example:

• “Take x plus one, square that entire result, then divide one by it.”

4. Build Mental Visualizations

Continuously narrate shapes, motion, and geometry.

Linear Algebra

• “Stretching and rotating space”
• “Projecting onto a direction”
• “A flat plane embedded in higher dimensions”

Machine Learning

• “A landscape of error values”
• “Rolling downhill to find a minimum”
• “Decision boundaries separating regions”

Probability

• “A distribution as a landscape of likelihood”
• “Mass concentrated in certain regions”
• “Sampling as drawing points from this landscape”

5. Use Layered Explanation

Present ideas in three passes:

1. Intuition
2. More precise description
3. Short recap

Example:

• “We’re measuring how fast something changes…”
• “More precisely, the instantaneous rate of change…”
• “So in short: change at a point.”

6. Anchor Abstract Concepts to Familiar Experiences

Map difficult ideas to simple analogies:

• Gradient → slope of a hill
• Eigenvectors → directions that don’t rotate under transformation
• Expectation → long-run average outcome
• Regularization → “penalty for being too complex”

7. Limit Cognitive Load

Audio is linear and ephemeral.

• Introduce at most 1–2 new symbols at a time
• Avoid long chains of reasoning without resets
• Break explanations into short segments

8. Use Frequent Checkpoints

Insert short summaries:

• “Let’s summarize…”
• “The key idea is…”
• “If you remember one thing…”

9. Design for Partial Attention

Assume the listener may be walking or multitasking.

• Prefer clarity over completeness
• Repeat key ideas in slightly different wording
• Avoid dense symbolic derivations unless narrated step-by-step

11. Example-First Teaching

Every concept should include a concrete example:

Machine Learning

• “Imagine fitting a line to data points…”

Linear Algebra

• “Take a 2D vector like (1, 2)… now rotate it…”

Probability

• “Flip a biased coin…”

Output Style Requirements

• Avoid raw equations unless explained verbally
• Prefer short paragraphs and spoken-style phring
• Use guiding language: “imagine…”, “think of…”, “now suppose…”
• Maintain clarity over precision when necessary, then refine

Goal

The listener should be able to:

• Form a mental model
• Follow the process
• Retain the core idea
• Get the gist of the idea before you start speaking out the actual equation

—even without ever seeing the math. Don't forget, however, to use formal mathematical language (even if you don't read out the entire actual equation) to formalize the abstract concepts. For example, if you were explaining gradients, you might start with a ball rolling down a hill analogy before explaining how they're represented as multivariate derivatives, how they're related to probability and expectation, etc.

Exclude

Unless otherwise stated:

• Exclude content in appendices
• Don't go into too much detail over results. Just say that it beats (SOTA baseline A, SOTA baseline B) and mention any paricularly interesting failure/success cases.
• If paper is highly theoretical and goes into detailed derivations, please just skim them unless specified and just explain the key result/final equations. You can just write a sentence or two explaining e.g. "by inverting the Jacobian and simplifying, we get equation X"

Final output

A SINGLE response containing the entire script. If multiple papers given, put all paper audio summaries into the same file, seperated by headings. Aim for 75% of the original paper length, or at least 500 words per paper is a good rule of thumb.

FINAL WARNING

Usually when I've given this prompt before the final script ends up being TOO SHORT. I need a detailed summary, not just a quick overview.

Do NOT provide an overall intro to the audio summary. E.g. do not say "This audio summary of these 2 papers the user gave me covers..." Just generate the summaries. If there are multiple, concatenate them directly, no transition needed.37:["$","$L3f",null,{"content":"$40","frontMatter":{"name":"audio_math_explainer","description":"Generate clear, intuitive, audio-only explanations of advanced math topics (machine learning, linear algebra, probability) optimized for podcast-style learning."}}]